Controllable Neural Machine Translation using Hints from Bilingual Dictionaries

The use of neural networks has allowed for significant advances in natural language processing in general and machine translation in particular. However, most widespread neural machine translation (NMT) methods do not provide control over words used in the output translation. Methods with such functionality could utilize hints from an external source to resolve ambiguities that inevitably arise when translating sentences from one language to another. This work introduces a novel method that uses bilingual dictionaries to align input and output words in the training translation examples. These alignments are used to hint which specific translation is wanted for an input word and train the model to indicate when decoding constraints should be applied. A described empirical study aims to demonstrate that, unlike prior methods, the approach has all three crucial qualities: high level of control over the generated text, little to no loss in quality compared to unconstrained translation, and no additional computational cost during translation.